JPH03155600A - Identification of word string in speech signal - Google Patents

Abstract

PURPOSE: To output a second best similar different sentence as well and to improve the reliability of recognition by providing the plural pieces of storage positions respectively provided with the N pieces of sub positions in a third memory. CONSTITUTION: A difference value is formed by comparing test signals with the reference signals of the plural pieces of prescribed words stored inside a first memory and the sum of the difference values is stored inside a second memory. In this case, the third memory for storing the pointer of a point where the word starts is provided with the plural pieces of the storage positions respectively provided with at least N pieces of the sub positions and the respective sub positions are provided with a first position for the address of the third memory, a second position for the address of the sub position within the storage position, a third position for word display and a forth position for the display of a difference sum. Then, by the contents of all the sub positions of the storage position inputted in the final test signal of speech signals, the respective kinds of different word strings are decided. Thus, many word strings provided with second best similarity to the speech signals are outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention is a method for recognizing at least one word string in a speech signal, comprising deriving from the word string a test signal representing a continuous time interval;
These test signals are compared with a plurality of predetermined words of reference signals stored in a first memory to form difference values, these difference values are added together, and the difference values are stored in a second memory. The sum is stored together with a pointer to a memory address such that the sequence of difference sums thus obtained can start at the beginning of a word and, at least at a word boundary, with a pointer to the word just ended. storing in a third memory a pointer to the point at which the word starts, and in at least one word string determined at the end of the speech signal, starting at least from the word for which the smallest sum of differences was obtained; A method for recognizing a word string in a speech signal, and a method for recognizing a word string in a speech signal, in which a word string is stored in the third memory, passing the start point of the word stored at the time and reaching the start point of the word by a pointer to the previous word. related to a device for

[Conventional technology]

Such a method is described in the Federal Republic of Germany Patent Application No. DE
- known from specification 033215868.

In this known method, a speech signal is compared with different words by using dynamic time adaptation,
Thereby, in the path of the speech signal in the recognition process a number of parallel word strings with similarities to the speech signal are obtained, which similarities are noted by the cumulative difference sums in the word strings. Finally, multiple word strings are terminated on the final speech signal, and the word string with the smallest cumulative difference sum is output as the only recognized word string.

However, as a result of different pronunciations, for example as a result of partial suppression of the ends of words, the word string thus obtained is not necessarily the string that corresponds to the uttered speech signal.

Therefore, in order to improve recognition, it has been devised to use a speech model that restricts the selection of the word(s) following the word just finished, in line with the conventions of natural speech. - Although this makes it possible to improve the reliability of recognition, each time as a result of similar-sounding words in a sequence of words that follow the conventions of natural speech, very similar but uttered sentences A word sequence that is not an accurate representation may be output as a recognized sentence, while a word sequence that reaches a slightly larger cumulative difference sum at the end of the speech signal may actually be a correct sentence. , it is still ultimately not rare. Therefore, in many cases, not only the word sequence or sentence with the best similarity is output, but also another sentence with the second-to-best similarity, especially if the best word sequence turns out to be inaccurate. It is effective to do so.

A judgment of likely inaccuracy is based on a knowledge source that, for example, is too complex and must be ignored in the cognitive process.

This is not easily possible using known methods, since only one word string is stored for all of the compared words at the end of the speech signal, and therefore This is because word strings that differ only slightly in similarity and end with the same word cannot be determined to be different.

[Disclosure of the invention]

It is therefore an object of the present invention to combine a large number of word strings with next-to-last similarity to a speech signal with no restriction on the individual word sequences other than that the individual word sequences differ with respect to at least one word. The solution is to adopt a method of the type defined in the opening paragraph, in such a way that the

According to the invention, the third memory has a plurality of storage locations each having at least N sub-locations, since this purpose is to recognize the N different word strings that most closely resemble the speech signal. and each of these sub-positions is
a first position for the address of the memory, a second position for the address of a sub-location within the memory location, a third position for displaying a word, and a fourth position for displaying the sum of differences; The address in the two positions represents a pointer to the start of a word (so that for each group of words, where the last word reaches the end of the word for the test signal, a new address a storage location, store this address as the starting point of each possible subsequent word in a second memory on the first reference signal of that word, and the information written to this sub-location will be stored in the same word. deriving from a memory location in which an address is stored in a second memory for a memory location of a first word that belongs to the group and reaches the end point simultaneously with respect to the final test signal; , and increment the sum of differences by comparison with the reference signal of the related first word and use only the smallest one. Contains the word
continues until all sub-locations of the new storage location are filled, and when deriving information, sub-locations of the address of the storage location from which information is to be derived from the sub-location. Write the address of the position in the first position, write the pointer of the sub-position from which the information is derived in the second position, write the just-ended first word of the association in the third position, and write the incremented sum of differences in the fourth position. and from the contents of all sub-locations of the memory locations entered during the final test signal of the speech signal, determine various different word strings and, through the display of the word in the third position, This is achieved by outputting the addresses of the storage locations included in the first and second positions of the sub-locations, the contents of these sub-locations, etc.

Proceedings of an academic conference held in New York in 1988''
Proc, IEEE Int, Conf, on
Acoustics, 5peechand Sign
al Processing”, New York
1988, pages 410-413, an algorithm for coherent speech recognition is known, which determines not only the best word sequence but also the word sequence with the next-to-last similarity. However, this requires a different recognition principle, namely multi-stage (num t-stag)
Method e) has been used, but no literature has been published regarding its specific technical implementation, especially the allocation of storage locations.

In the method of the invention, the contents of the third memory are enlarged in a certain way so that it is now possible to form a large number of different word sequences, each time these different word sequences with the best similarity to the speech signal. that only the following will be considered further:
A step for the generation of information for a new storage location in the third memory is guaranteed. Furthermore, the present invention provides that the end of a word cannot be followed by any other word of the whole word, but the end of a word of a certain group, preferably of a certain synclass class. This allows the use of a speech model in which only the word just read or the word commanded by the starting point of this word can be followed.As is well known, the reliability of this word recognition increases significantly. make it possible.

In each storage location of the third memory, the difference sum itself can be stored in the fourth position of each sub-location. Another possibility according to embodiments of the invention, which requires less calculation time, is that the absolute value of the difference sum is stored in the fourth position of the first sub-location of each storage location, The difference between the difference sums and the difference sum of the first sub-position are characterized in that they are stored in the fourth position of each subsequent sub-position. In this way, especially if a particular test signal results in only one word, the difference between the difference sums remains valid and therefore only the absolute value of the first sub-position has to be increased. This is achieved.

If a large number of words end simultaneously for the same test signal, in order to obtain information for new storage locations in the third memory, the sublocations of the storage locations corresponding to these words must be the sublocations with the smallest sum of differences. must be mixed in such a way that only the ingredients are ultimately used. An efficient way of carrying out this mixing process in another embodiment of the invention is that for one of the words arriving at the end for the same test signal, the address is stored together with the sum of the differences of that word. Further information is obtained from the information of the memory location in which it is located and is stored in a memory location with a new address in a third memory, and for each sub-location the corresponding of every other one of these words is determined. The information of the memory location to be compared is sequentially compared with the information of all sub-locations of the new memory location, and if the information of the two mutually compared sub-locations indicate the same word sequence previously examined, The larger difference sum is suppressed, and the unsuppressed information of the just compared sub-locations of a word is added to the two new storage locations whose difference sum is greater or less than the difference sum of the compared sub-locations. inserted between the sub-locations, and the sub-location information of the new storage location is 1 if necessary.
It is characterized in that only the sub-position is shifted.

In this way, there is no need to iteratively search through each individual sub-position of a terminating word; the process may begin with any of a number of simultaneously terminating words, and information about the starting word is not required. 3 to enter from
The data at the new address of the th memory is obtained, and the information of other words that end at the same time are loaded into the new memory location, and the data that was previously in this new memory location is erased or shifted out and finally All words of the same group ending at the same time end up being processed. This minimizes the time required to combine data from multiple simultaneously terminating words.

The preceding word sequences that have been examined in detail are compared with the two
In order to check whether the information from the two sub-locations is different, the chain of third memory locations following the data in the first and second sub-locations must be examined each time. In yet another embodiment of the invention, this is achieved in the fourth memory, each time the information is written to a sub-location of the third memory, by a word that has been previously examined and just passed. The instruction of the stretched word string is stored in place of the instruction of the finished word at the new address stored in the sub-position inserted in the third position, and the previously examined word string is It is determined via the address of the fourth memory, which can be simplified in that the information about the sub-location to be inserted is stored in the sub-location obtained therefrom.

This requires yet another memory, a fourth memory, but only the corresponding inputs to this memory have to be considered to check whether the two compared word strings are similar. . This saves a considerable amount of processing time. Furthermore, at the end of the speech signal the word string with the best similarity is
The word string can be read directly from this fourth memory without having to determine it by tracing back the criteria or instructions in the first and second sub-locations.

Apparatus for carrying out the method according to the invention, comprising: a speech signal processing device for obtaining a characteristic test signal; a first memory for storing a reference signal for the word to be recognized; and a difference value. a comparator circuit for comparing each test signal with a reference signal and an indication of the difference sum and the beginning of the column of difference sums for the corresponding word in order to a second memory for storing, and a third memory for storing, when the end of the word is reached, a pointer to the beginning of the sequence of difference sums and a pointer to the word just passed. When the end of the word is reached, for each newly addressed storage location, the third memory has a number of sub-locations, each of the sub-locations having four storage positions; A processing circuit is provided which addresses the same storage location in the third memory for all words belonging to the same group of words and ending with the same test signal and stores them into separate sub-locations. Write the information obtained from the read contents of the sub-location of the location, the storage address of that information being the one indicated by the input corresponding to the finished word in the second memory, and the sub-location as described below. i.e., for the sub-position for which the sum of differences is stored and where the sum of differences increased by the increase in the sum of differences as a result of the comparison of the reference signals of the corresponding first word is the smallest, and , the processing circuit is characterized in that the processing circuit obtains information only from sub-positions for which the sequence of preceding words containing the instantaneous words examined up to that point differs.

The drawings will be explained below.

Test signals are generally taken at regular intervals from the speech signal to be examined, for example at intervals of roms to 20 m5.

These test signals can be, for example, the short-term spectrum of the speech signal, the fundamental speech frequency, the loudness or similar values that can be prepared especially for word recognition. Methods for generating such test signals are known and are outside the scope of the present invention.

It is known that speech signals consist of individual words derived from vocabulary. A draft word corresponds to a set of columns of reference signals derived from individually pronounced words. Let the reference signal, and therefore the sequence of words, be k·l,−
-- Indicated by. Each reference signal in the column has a sign j=1.
---j(k) was attached. Here, j(k) indicates the length of the reference signal string.

The ultimate goal of recognizing words in a coherent word string is to determine the consecutive plurality of reference signal sequences that best adapt to the test signal sequence derived from the speech signal.

The aforementioned Federal Republic of Germany Patent Application No. DE-O33
A method for determining a single word sequence best adapted to a speech signal, known from German Patent No. 215,868, is explained in detail with reference to FIG. A test signal i of a speech signal and a reference signal j in the individual columns (here five columns 1 to 5 of five words each are shown by way of example) form a matrix 10 of matrix points (1+j,k). stipulates. A local difference value d(L L k) is assigned to each matrix point and serves as a measure of the deviation from the corresponding acoustic characteristic or the difference between these acoustic characteristics. The problem in recognizing coherent word strings is transformed into the problem of finding a path through a matrix of matrix points (i, jt k) that represents the best adaptation between a test signal and an unknown succession of reference signal sequences. Can be done. In other words, the sum of the difference values d(+, Jl k), when increased by a time distortion value that depends on the direction of the previous matrix point of association, becomes maximum for all the matrix points on the above path. Become. Starting from this optimum path, the succession of the reference signal train can be determined unambiguously, as is clear from FIG.

The optimal path is determined by non-linear adaptation of the test signal train to the respective reference signal train. For this purpose, we use the fact that the optimal path through a matrix point (i, j, k) consists in part of the optimal subpaths preceding this matrix point. For this purpose, we use a matrix point (1,
The minimum difference sum D(i,
j, k). Since this difference sum is a sum of local difference values, it can be easily converted into the sum of the difference value along the optimal path to the preceding point and the local difference value of matrix point (i, j, k). Decided. Next, for the optimal path, it is necessary to determine the leading point with the smallest sum of differences. In the known method, doing so results in the following transition rules for transitions within a train of reference signals, without allowing for time distortion values:

D(i, j, k) = d(i, j, k) + min D(i-1, j, k), D(i-1,
j-1,k), D(i, j-1,k) Since the optimal path is not yet known, the above equation yields multiple paths.

Among these passages, FIG. 1 shows, in addition to passage P1, which is finally found to be the optimum passage, passages P2 and P
3 is also shown.

If we do not use a threshold for the difference sum, we actually get a separate path for each reference value of all the difference signal sequences;
These paths do not intersect, especially at word boundaries, according to Bellman's optimality principle.

For a test signal, in order to determine the difference sum with respect to all reference signals, only a small part of the total matrix of difference sums of the preceding points is used, i.e., the preceding test signal i Only the difference sum associated with -1 is needed. These difference sums are stored in a memory, diagrammatically shown in block 12, in section 12a and are rewritten on each new test signal.

Furthermore, it is possible to trace back the optimal path, i.e. each transition position from one column of the reference signal to another,
That is, it is necessary to trace back points of paths located at word boundaries. Since the ultimate goal of conventional methods is to determine unknown sequences of words or sequences of reference signals in a speech signal, for which test signals a subpath is initiated that ends at the end of the sequence of reference signals. It is appropriate to decide whether to do so. However, for word recognition in the load string, the subpath details within the reference signal train are not appropriate.

Initially, the end point of the subpath of the optimal path at the end of each column of the reference signal is not yet known, and for this reason it is necessary to keep traceback information for the entire duration of the time-adaptive processing. The optimal path is the matrix point (i, j, k)
Each reference signal sequence has its own starting point for the first reference signal j·1 in this reference signal sequence. Therefore, for each matrix point a traceback pointer B (b Jl k) can be determined as the test signal address that yields the best path to this matrix point (L L k).

Therefore, one column of the traceback pointer B(j, k) needs to be stored each time, as for the difference sum. That is, one storage location 12a for the difference sum and one storage location 12b for the traceback pointer is required for each reference signal j of each column in the reference signal string. This is illustrated in the left part of FIG.

However, for traceback, it is not necessary to determine the start of the subpath in the associated reference signal sequence, but rather the end of the preceding subpath immediately preceding the start of the final subpath, as described above. .

Therefore, the conventional method is simplified by directly storing the preceding ending test signal address in the traceback pointer B(j, k) instead of the starting test signal address.

As previously mentioned, only the traceback pointer at the end point, i.e. the final reference signal J(k) for each reference signal sequence, allows the sequence of words to be traced back along the optimal path. . The reason is that the traceback pointer B(J(k), k) at the end point
This is because it defines the end point of all the columns that precede it. However, since the traceback pointers at these end points stored in the memory 12 are rewritten during the processing of the next test signal, a separate memory is required to store the traceback pointers at the end points of each column of the reference signal. becomes.

Since such an end point can occur for each test signal i, i.e. at least one end point can occur regardless of the speech model due to the optimality principle described above, this separate memory stores a traceback pointer for each test signal i. It is necessary to have a storage location for . This separate memory is diagrammatically shown in FIG. 1 as a block 14 below the matrix 10, with a traceback pointer B(J(k),k) in a section 14a of this memory.
is always stored in the form of the test signal address F(i) of the preceding end point where the difference sum D(J(k), k) is the minimum.

However, since the test signal addresses F(i) are not the end points themselves of the individual sub-paths of the optimal path within each reference signal train, but are words associated with the associated reference signal train, the reference signal strings representing these words are Also the column number of the associated initial address F
It is necessary to memorize it together with (i). We denote the stored column numbers by T(i), and these are the ones with the smallest difference sum D(J(k), k) at the end point J(k) for the same test signal with respect to all other reference signal columns. This is the column number of the reference signal column that appears.

As is clear from FIG. 1, the test signal address 1 (L-1) at the end point of the preceding reference signal train and the row number k (L) of the final reference signal train are stored in the final test signal engineer. can be extracted from the given value. Similarly, the other test signal addresses F(i) stored are storage locations 1(t-1) and 1(L-2) of the memory 14.

1 (L-3), etc., the column numbers k (t-1), k (t-2) of each preceding reference signal sequence
--- and the test signal addresses 1(L-2), 1(L-3)--1 of the end points of the reference signal sequence ending before it directly yield the stored addresses.

In this method, when using a speech model in which two or more reference signal sequences that can be continued only with other specific reference signal sequences depending on the speech model can be terminated at the same time for the same test signal, some In this case, it is not possible to simply store or proceed with the end of the reference signal sequence with the smallest difference sum, but with a large difference sum. Other reference signals ending at can ultimately yield small difference sums at this end point in that they are later better adapted to the speech signal.

However, this problem can be easily solved by addressing the memory 14 separately in FIG. need to remember,
These addresses must be completely derived from the instantaneous current column number of the test signal.

However, not only the word sequence that best approximates the speech signal, i.e., has the smallest sum of differences at the end, but also the sequence that is the next best similar to the speech signal, i.e., has the next largest sum of differences. A bigger problem arises when word sequences also need to be determined and read out. The reason is that it represents a word string with sequentially large difference sums,
Therefore, multiple words of the same syntactic group may appear in the same test signal, which must be remembered at the end until it is ascertained which is the N best sequence of words that actually best resembles the speech signal. This is because there is a risk that it will be terminated.

〔Example〕

FIG. 2a shows the composition of the memory corresponding to the memo January 4 of FIG. This figure diagrammatically shows a number of memory locations 31 to 37 for the traceback pointer of the memory, each of these locations having three multi-position sub-locations for determining the three best word sequences. . Figure 2b shows one storage location in more detail. In this embodiment, each storage location has three sub-locations TP1.
．． TP2. TP3, each of which has four storage positions Stl.

Si2. Si2. I have a Si4. This allows determining the three best word sequences.

If the next-to-last word sequence also has to be determined, the number of sub-positions must increase accordingly, but the number of positions remains the same. The information in the first sub-position TPI represents the best word sequence that most closely resembles the speech signal, and the sub-position TP2 identifies the second-to-best word sequence with the next largest sum of differences, as follows: The same is true.

Adjacent similar storage locations of this memory that are eventually filled in the case of adjacent test signals or in the case of words of other synclass groups ending in the same test signal obviously have significantly larger or significantly smaller difference sums. word sequences, since the optimal sequence cannot be determined until the end of the speech signal is reached.

The contents of the individual positions are as follows: Position Stt contains the address L of the previous storage location of this memory following the word just finished. Position St2 is a sub-position n whose address is stored in position Stl.
, and then the information of that sub-location is obtained in the manner described below.

For the position St3, it is assumed that the reference signal, ie, the column number k of the immediately ending word string, is stored. Finally, position St4 contains the difference sum arrived at for the word just finished and the word string preceding this word.

In FIG. 2a, storage locations 31 to 37 have a further storage position shown at the top right corner of each block representing storage locations, which corresponds to all sub-locations of this storage location St4 of each sub-location. The minimum difference sum is stored and thereafter includes only the difference between the related difference sums and the minimum difference sum. However, in this case, since the sub-position TPI implicitly includes the minimum difference sum, the position St4 of the first sub-position TPI inherently contains the value 0, so the position St4 of the first sub-position TPI is It can be used to store the absolute value of the minimum difference sum of this storage position, while position St4 of the second sub-position TP2 and thereafter can be used to store the absolute value of the minimum difference sum of the second sub-position TP2 and the above-mentioned minimum difference sum. including the difference d. Therefore, the further storage positions for the minimum difference sums shown in blocks 31 to 37 of FIG. 2a are not actually required.

The process of tracing back the word string preceding the end word, which could also be the end of the speech signal, is illustrated by means of the arrows in FIG. 2a. That is, the word string found to be the next best based on the sum of differences in memory location 37 is read so that the column number k of the last ending word is the second best word string in memory location 37 in the second row. 2 based on the corresponding instruction n in the second position of the second sub-location of memory location 35, i.e. the second sub-location of memory location 37, read from position 3 and using the address L in the first position of the second row.
The th sub-location is addressed and the column number of the preceding word stored in the third position of the second sub-location of memory location 35 can be read. Furthermore, using the address in the first position of the second sub-location of memory location 35, memory location 34, i.e. the first sub-location of memory location 34, is updated to the corresponding information in the second position of memory location 35. The path is then further directed towards the beginning, i.e. memory location 32.
is traced back to the first sub-position of , where the first word of the second-best word string ends. In fact, further locations belonging to other word strings are located between locations 30 to 37, of which only locations 31, 33, and 36 are illustrated.

The word string found to have the minimum difference sum at memory location 37 is stored in memory location 36, 35, 33, for example as described above.
．． 31, i.e. each time via the first sub-location, because the speech signal only travels through equivalent or higher-order sub-locations - once the second storage location. This is because a word string extended through the sub-locations of can no longer extend through the first sub-location of the memory location.

How each new storage location information is generated or obtained in accordance with the storage location information of FIG. 2a is explained in Section 3.
This is explained with reference to FIG.

It is based on a speech,model in the form of a graph with nodes and intervening,links, where a node can be seen as a common point for all,links following a similar path. FIG. 3 shows a portion of such a speech model, which has three nodes N1, N2, N3 and two words, 4, k5, according to the speech model idiom, two links towards node N1. . It is more appropriate to use the term "link", since one and the same word can be represented by two or more different links, i.e. links are unique but different within the speech model. This is because you can return to the same position.

Similarly, link 6 and k7 go to node N2.

In fact, the number of links pointing to the majority node increases as the allowed word possibilities expand.

Furthermore, another link is provided from each node, which means that, for example, a 4 link in the speech model can only be continuous with a specific other link, and a 1 link is shown in more detail in Figure 3. has been done. This can also be said about 5 links.

Similarly, in this case, link 6 and link 7 can be continued only with link 2 and link R3, for example. The two nodes N1 and N2 do not necessarily have to be reached by the corresponding links at the same time (also due to the dynamic time adaptation applied and the consequent increase in the number of paths along each individual link). It should be borne in mind that , different and generally successive test signals will repeatedly reach the end of this link.

Link 1, Link 2 and Link 3 are now nodes N
This will be followed by the word leading to 3, i.e. corresponding to link 8 to link 9. Depending on the speech model, after node Nl there will be one more node. It is also possible that the part of the sentence that can be reached by another link, i.e. that has already extended to node N1, can be continued in two or more different grammatical paths, but this is not the case in the speech model. This is a matter of detail and will not be discussed further at present.In principle, it is also possible to use the method described without using a speech model, i.e. in that case the speech model consists of only one node. , and all links established from this node return to this node.

FIG. 3 does not give any timing conditions, ie, it does not indicate when and for which test signals the transitions arrive at the nodes. However, the link k
l, 2 on the links shall arrive at node N3 at the same time, ie for the same test signal. If you do this, the first
In the matrix corresponding to the figure, several paths are combined to one point, and these paths represent word sequences that have hitherto been elaborated with a different similarity to this real speech signal, some of which are continuous They require traceback from the end to be possible. The number of routes to be continued is node N3
equal to the number of paths along each link terminating in , i.e. the total number of paths must decrease. This reduction while preserving the possibility of traceback is explained in more detail using FIG. As a result, this is done in particular in the memory for the respective initial address of the individual column, i.e. the memory corresponding to memory 14 in FIG. Involved. Each link that is specifically assigned to one word and therefore to one reference signal sequence is stored in the memory 14.
The address of the storage location within is communicated using the traceback pointer B(j,k) formed at the beginning of this link and stored in section 12b of note January 2 of FIG. For link kl this is address L1 and for link 2 this is address L2.

A new storage location in memory 14 with address L3 is read at the end of Link 1 and Link 2. The simplest way to accomplish this is to select the next free storage location.

The content of memory location 3, ie the information written therein, is obtained sequentially from memory location 1, the address of which is read from memory 12b at the end of link kl, as already mentioned. The first position of all sub-locations of this memory location tl contains the associated preceding address, and the address t
a and address Lb are shown individually as examples. The contents of the second position indicate from which sub-location with address ta and address Lb the corresponding value was obtained, that is, in this example, it was obtained from the first sub-location of both storage locations. ing. As described below, the order of the associated links from which the sub-position information originates is stored in the third position, i.e.
5 results in a smaller difference sum S', while 5 results in a larger difference sum by the value d'l. This address tl is transmitted to all links emanating from node Nl for the same test signal, but in this case only link kl is considered. The sum of differences S' is node N
of all links emanating from 1, and consequently of the starting link kl, is used as the value D(J, K), i.e. before comparing the first reference signal of 1 to the link with the instantaneous test signal. and during the path of the link, it increases according to a dynamic time adaptation between subsequent comparisons of the test signal and the reference signal.

At the end of link kl, the associated quantity D(J(kl
).

k) reaches the value S1, which is written in the fourth position of the first sub-position in the storage location 3.

The third position of this sub-position stores 1 in the order,
The second position, on the other hand, stores the number of sub-locations of the previous storage location tl from which information is obtained, ie the value l. Finally, the address tl enters the first position of the first sub-position. The second and third sub-locations are loaded in a similar manner so that the information initially in the second sub-location in the example of FIG.
In the second sub-location, the information from 2 to the second simultaneously ending link, i.e. the information from the storage location with the address L2 stored with this link, and therefore also the information from the first sub-location, to the link 2. This is because the difference sum S' increased by the difference value along the line is smaller than the difference sum that originally existed at the second sub-location of storage location 1, and therefore was subsequently inserted. Then, the absolute difference sum at the end of link 2 is the difference sum Sl at the first sub-position.
is submitted to yield a difference d1, while storage location t
The difference value d' 1 from the second sub-position of l corresponds directly to the difference value d2 stored during the last position of the third sub-position of storage position 3. The content at the beginning of the third sub-location of memory location 3 that originated from the third sub-location of memory location 1 is now shifted out by the input from memory location 3 because it is too large. This is because it represents the sum of differences. Similarly, information for storage location 3 is obtained from the second and third sub-locations of storage location 2.

Address L3 is traceback pointer B (1,
k) in the links emanating from the node N3 together with the first criterion value of each of 8 or 9, and thus transferred, the corresponding value D(1,k) is therefore the difference sum S1 and each link. is formed from a comparison with a first reference signal.

When a test signal is being considered, if at least one other link simultaneously terminates to a node other than node N3, another storage location must be provided, for example with address L4. , the information to be input thereto is obtained in the same way as described above.

- To be successful, a storage location must be provided for each node where at least one link terminates for the same test signal.

However, before the information from the first sub-location of storage location 2 is loaded into the second sub-location of storage location 3, different word sequences of different affinities for the speech signal must eventually be determined. Therefore, starting appropriately from the last memory location of memory location 3, it must be ascertained whether this information does not relate to the word string that reached memory location 3 via memory location 1. If the information to be loaded into a sub-location represents a word sequence that is already present in one of the sub-locations of this storage location, this information should be suppressed, i.e. this information is larger It represents the absolute sum of differences. If this becomes a sub-location of memory location 3 that has been emptied, if this can be done based on the sum of the differences between the other sub-positions and the sub-position to be inserted, It would be possible to write the information to be inserted. In this case, the already possible sub-locations within storage location 3 do not need to be shifted if a new sub-location is to be inserted.

In order to check whether a word string corresponding to the word string to be inserted already exists in memory location L3, a link issue from memory location L3 is inserted into the third position of one of the sub-locations in memory location L3. It must first be checked whether it already exists, and in that case the corresponding destination columns of the link determine whether these columns of the link differ by at least -position or match up to the beginning. must be traced back to do so.

This rather complicated retracing can be avoided if an additional word string memory address is stored in the third position of each sub-location instead of in the order of the last ending link, and that memory Within, if this memory does not already contain this word string, it is traversed to the link and stored in a new storage location for every end of link information for the link string containing the link. This means that at the end of the speech signal, the individually determined word strings of the third position in the last storage location of the memory 14 read at that time coincide with their increasing similarity with the speech signal. can be directly read out via the additional word string memory.

A possible organization of this word string memory will now be explained in some more detail with reference to FIGS. 5 and 6.

As an example, FIG. 5 gives some word sequences that may occur in the process of recognizing a speech signal, but for simplicity only a few word sequences are traced. Yen W1～
W2 represents the end of an individual word and the contents represent the address of the word string memory illustrated in FIG.

Starting from the starting point WO, where no word has ever ended, we begin with two words above all: “who” and “when.”
'' is compared with the speech signal based on a predetermined speech model, for example. When the word "who" ends, it is registered at address W1 of the word string memory shown in FIG. Section El with back pointer and section E2 with link “who”
and. Section E3 remains empty at that time. For example, if the word "when" ends at a slightly different moment, i.e. if another entry is made in the word string memory of FIG. 6 at address W2 for another test signal, then in section El. The back pointer of also points to the starting point wO, and section E2 contains the link "when".

Now, the comparison is especially between the words “has” and “WaS”.
It is made about. The word “ha” generated from point W1
If ``s'' is the next word to reach the end, a new entry is made in the word string memory at address W3, and that memory has a back pointer to wl in section El and a link ``has'' in section E2. At the same time, the link "has" is entered in section E3 at address wl as a link originating from point W1.

If the word "WaS" subsequently originates from point W1, it is checked whether this link already exists in section E3 at this address W1. Assuming this is not the case yet, a registration is made in the word column memory at address W4, and that registration has a back pointer to Wl in section El and a link in section E2.
Furthermore, the registration in section E3 at address Wl is completed by this link "WaS".

Similarly, the word “WaS” generated from point tV2
” and “has” are completed and the corresponding word is also completed in section E3 at address W2, then
New addresses W5 and ~v6 are created.

Now, above all, the words "written" and "mai"
"led" are compared starting from point ~v3. For simplicity, the words actually following W4-W6 are not considered.

When the word "writen" is finished, a new registration W7 is made in the word string memory, which registers a back pointer to W3 in section El and a back pointer to section E2.
It includes a link "written" within. This word is also completed in section E3 at address W3. The same thing is written to the other word "mailed" written to address W8, and also to address ~V9. W
I.O.

It also applies to other words that generate registration in Wit and Wl. All section E at the destination address
3 is completed in the same way.

Now, starting from the point Wl, for example, if the word “has ended at the latest moment via another path, we immediately know that this word sequence already exists via the initial address W1 from the register E3. Is it possible to be sure?

At the end of the speech signal, from the last storage location in the third memory 14, the last word at the same time as the end of the word string and other previous words through it, is transferred to the third memory location of all sub-locations of the word string memory. It is now possible to determine each time by the second section E2 of the address contained in the position and to read these out as a word sequence.

In the block diagram shown in FIG. 7, the speech signal to be recognized is applied to a microphone 30 and converted into an electrical signal that is supplied to a speech signal processing circuit 42. In the block diagram shown in FIG.

In this circuit a characteristic test signal is obtained from the speech signal - since real-time processing is not possible in the case of large power supplies for fishing boats, for example short consecutive time intervals of Ioms in several adjacent spectral regions. The amplitudes of, for example, speech signals and appropriate test signals are temporarily stored. This intermediate memory, not shown, is addressed by an address generator 44.

Subsequently, the test signal is passed through the connection line 43 to the comparator circuit 16.
The circuit simultaneously receives and receives the reference signal from the reference memory. This reference memory 18 is connected to the address generator 2
4, the address generator reads out all the reference signals in succession or, if a threshold is used, reads out a particular reference signal of the reference signals stored in the reference memory 18. Once all relevant reference values are applied to the comparator circuit 16 and the address generator 24
supplies a signal to the address generator 44 via the connection line 25, so that the next test signal is applied to the comparator circuit 16.

A comparator circuit compares each test signal with an applied reference signal and forms a difference value for each applied reference signal, and for example as described in German Patent Application No. DE-O332.
According to the rules of dynamic programming as known from the publication No. 15868, a difference sum is obtained from these values and stored in the memory 12 which is also addressed by the address generator 24. Furthermore, a back pointer is stored in this memory at each address to indicate at which position or for which signal the sequence of differential signals traversing the associated word began within this word. .

At the end of each word, i.e. the address generator 24
addresses the last reference signal j=J(k) in a word, the processing circuit 20 is activated and it retrieves the back pointer B stored at this address from the memory 12 via the connection line 13. Read and drive the memory location in the note 1/4 at the address that corresponds to the back pointer to read this memory location via the connection line 21. The just finished word is the first finished word in the associated test signal, or if a speech model is used, the processing circuit 2 by the speech model memory 14.
the first of the grammatically contiguous set of words determined by 0
If it is a word, the next free storage location in the memory 14 is addressed via the connection line 23 and the contents of the read storage location are stored therein. however,
If the word just finished is not the first word for this test signal, the contents of all sub-locations of the read memory location of note January 4 are filled by the word contained therein and momentarily passed. Checks to determine whether the increased difference sum is less than the difference sum of the sub-locations of the memory location in memory 14, which memory is newly filled for this test signal for the first ending word, and The contents of this newly filled word are maintained for its memory, for example in the processing circuit 20. If such a sublocation is found, the content of that sublocation and the content of the next sublocation are advanced by one sublocation, the content of the last sublocation is lost, and the content of the sublocation to be read is The just compared contents of the positions are then written into the emptied sub-positions.

This is applied successively to all two sub-positions until a sub-position is found whose increased difference sum stored therein is greater than the difference sum of the last sub-position of the last memory location read. Ru. While reading a new sub-location, the contents of the individual storage locations are obviously updated in the manner described above, and the number k of the word just finished is obtained from the address generator 24.

Instead of reading into a new memory location in the memory 14 for the first ending word in the case of a new test signal, at least if a threshold value is used for the difference sum, if the comparison is allowed, the reference memory The entire contents of this storage location are initially stored intermediately in the processing circuit 20 until all reference signals of all 18 words have been compared with the instantaneous test signal, and the next can be completed by the end word of . The information obtained can then be loaded into a new storage location in memory 14. The essential feature is that the individual sub-locations of this storage location are arranged in columns of their difference sums.

The address of this newly read memory location in the memory 14 in each case is also applied via the connection 23 to the comparator circuit 16, which circuit stores this address as a pack pointer B for every new starting word. 1
Write into 2. In this method, the memory 14 is read successively until the last test signal of the speech signal to be recognized is compared, after which the recognized word sequence is read out by reading out the memory 14 as hereinabove described. It is output by the processing circuit 20 and applied to an output device 38, such as a printer or a memory or further processing circuitry, for further processing of the recognized text.

The comparator circuit 16 and the processing circuit 20 may be constituted by a jointly programmable computer, which computer then comprises, for example, address generators 24 and 44 and/or separate memories, in particular two memories 1.
2 and 14 (may be formed by the corresponding address portions of the common memory).

[Brief explanation of the drawing]

1 is a diagram illustrating a method for determining a word string; FIGS. 2a and 2b are illustrations illustrating the organization of the storage locations of the third memory and the traceback of the word string in the method according to the invention; FIG. Figure 3 is a diagram showing a part of the graph of the speech model; Figure 4 is an explanatory diagram showing a method of forming information on a new storage location in the third memory; Figure 5 is a typical diagram. FIG. 6 is an explanatory diagram showing the contents of the fourth memory corresponding to the combination of words shown in FIG. 5; FIG. 1 is a block diagram illustrating an apparatus for implementing the method according to the invention; FIG. IO...Matrix 12, 14...Memory 12a, 12b, 31-37...Storage location 14a
, 14b...Storage section 16...Comparison circuit 18...Reference memory 20...Processing circuits 24, 44...Address generator 38...Output device

Claims (1)

Claims: 1. A method for recognizing at least one word string in a speech signal, comprising deriving from the word string test signals representing continuous time intervals and storing these test signals in a first memory. forming a difference value by comparing the reference signals of a plurality of predetermined words, adding these difference values, storing the sum of these difference values together with a memory address pointer in a second memory; The pointer to the memory address is such that the sequence of difference sums thus obtained can start at the start of a word, and furthermore, at least at a word boundary, the pointer to the point at which the word just ended can be used to point to the point at which the word starts. at least one word string stored in memory and determined at the end of this speech signal starting at least from the word for which the smallest sum of differences was obtained and passing through the start of the word currently stored; and a method for recognizing a word string in a speech signal in which a word string leading to the start of the word from a pointer to the previous word is stored in this third memory. To recognize a string, the third memory has a plurality of storage locations each having at least N sub-locations, each of these sub-locations consisting of: a first position for an address of the third memory; and a storage location. It has a second position for the address of the sub-position within, a third position for displaying the word, and a fourth position for displaying the sum of differences, and the address in the first two positions is the starting point of the word. For each group of words whose last word reaches the end of the word for the test signal, a new storage location is addressed in the third memory, and this address is transferred to the second memory. The information written in this sub-position belongs to the same group of words and reaches the end point at the same time for the final test signal. deriving an address from a storage location stored in a second memory for the storage location of the first word stored in the first word; The sum of differences obtained by comparing a word with a reference signal is incremented and only the smallest one is used. , and continues until all sub-locations of the new storage location are filled, and when deriving information, the address of the sub-location of the address of the storage location from which information is to be derived is placed in the first position. Write the second pointer to the sub-position to derive the information.
position, write the just-finished associated first word to the third position, write the incremented difference sum to the fourth position, and all of the memory locations entered during the final test signal of the speech signal. From the contents of the sub-locations, various different word strings are determined and, through the representation of the word in the third position, are expressed as the addresses of the memory locations contained in the first and second positions of said sub-locations and A method for recognizing a word string in a speech signal, characterized in that the word string is output together with the contents of these sub-positions. 2. The absolute value of the difference sum is 4 of the first sub-position of each storage location.
4th position of each subsequent sub-position, and the difference between the difference sums at this sub-position and the difference sum of the first sub-position are stored in the fourth position of each subsequent sub-position. The method described in 1. 3. For one of the words reaching the end for the same test signal, further information is obtained from the information of the storage location where the address is stored together with the difference sum of that word, and this is added to the third , and for each sub-location, the information of the corresponding memory location of every other one of these words is sequentially combined with the information of all the sub-locations of the new memory location. are compared, and if the information of the two mutually compared sub-positions indicates the same word sequence examined previously, the larger difference sum is suppressed and the just-compared sub-position of a word is The unsuppressed information of a position is inserted between two sub-locations of a new storage location whose difference sum is greater or less than the difference sum of the compared sub-locations, and the information of the sub-locations of the new storage location is required. 3. A method as claimed in claim 1, characterized in that the method is shifted by one sub-position. 4. In the fourth memory, each time the above information is written to a sublocation of the third memory, the word string pointed to by the previously scrutinized word just finished is replaced by the word just finished. Instead of the instruction in the sub-location inserted in the third position, the word string stored in the new address stored in the inserted sub-location and previously examined is determined via the address of the fourth memory; 4. Method according to claim 1, characterized in that the address is stored in the sub-location from which information regarding the sub-location to be inserted is obtained. 5. Apparatus for carrying out the method according to any one of claims 1 to 4, comprising a speech signal processing device for obtaining a characteristic test signal and for performing recognition. a first memory for storing a reference signal for the word in which the test signals are stored, and a comparator for comparing each test signal with the reference signal to form a difference value and to cumulatively add the difference values to form a difference sum. a second memory for storing the difference sum for the word in question and an indication of the beginning of the series of difference sums, and when the end of the word is reached, a pointer to the beginning of the series of difference sums and a pointer to the beginning of the series of difference sums that has just ended; and a third memory for storing a pointer to a word that has been addressed. When the end of the word is reached, for each newly addressed memory location
4) has a number of sub-locations (TP1, TP2,...), each of which has four storage positions (l, n, k
, d), and a processing circuit (20) is provided, the processing circuit (20) having:
Address the same memory location (l1, l2, l3,...) in the third memory (14) for all words belonging to the same group of words and ending with the same test signal (i) and separate Write into the locations the information obtained from the read contents of the sub-locations of those storage locations, the storage address of which information is the input (B,
j, k), and for which sub-positions the difference sum is stored and as a result of the comparison of the reference signals of the corresponding first word. where the increased difference sum is the minimum due to an increase in the difference sum, and
Apparatus characterized in that the processing circuit (20) obtains information only from sub-positions for which the sequence of preceding words containing the instantaneous words examined up to that point differs. .